图像特征提取之Hog特征提取
HOG全称（histogram of oriented gradients）,方向梯度直方图，可以用来提取表示图像的特征，本质就是一行高维特征。
HOG特征提取步骤
-
图像预处理（gamma校正和灰度化）【option】
-
计算每一个像素点的梯度值，得到梯度图（尺寸与原图一致）
sobel计算水平和竖直梯度，并通过公式求得梯度的方向（边缘方向与梯度方向垂直）
梯度方向取绝对值，梯度方向取值范围为[0,180]
-
统计每个cell的梯度直方图（不同梯度的个数），形成每个cell的descriptor
-
将每几个cell组成一个block（3*3），一个block内所有cell的特征串联起来得到该block的HOG特征descripitor
-
将图像image内所有block的HOG特征descripitor串联起来得到该image的HOG特征descripitor，即最终分类的特征向量
HOG特征提取python实现
import cv2
import numpy as np
import math
import matplotlib.pyplot as plt
class Hog_descriptor():
def __init__(self, img, cell_size=16, bin_size=8):
self.img = img
self.img = np.sqrt(img / float(np.max(img)))
self.img = self.img * 255
self.cell_size = cell_size
self.bin_size = bin_size
self.angle_unit = 360 / self.bin_size
assert type(self.bin_size) == int, "bin_size should be integer,"
assert type(self.cell_size) == int, "cell_size should be integer,"
assert type(self.angle_unit) == int, "bin_size should be divisible by 360"
def extract(self):
height, width = self.img.shape
gradient_magnitude, gradient_angle = self.global_gradient()
gradient_magnitude = abs(gradient_magnitude)
cell_gradient_vector = np.zeros((height / self.cell_size, width / self.cell_size, self.bin_size))
for i in range(cell_gradient_vector.shape[0]):
for j in range(cell_gradient_vector.shape[1]):
cell_magnitude = gradient_magnitude[i * self.cell_size:(i + 1) * self.cell_size,
j * self.cell_size:(j + 1) * self.cell_size]
cell_angle = gradient_angle[i * self.cell_size:(i + 1) * self.cell_size,
j * self.cell_size:(j + 1) * self.cell_size]
cell_gradient_vector[i][j] = self.cell_gradient(cell_magnitude, cell_angle)
hog_image = self.render_gradient(np.zeros([height, width]), cell_gradient_vector)
hog_vector = []
for i in range(cell_gradient_vector.shape[0] - 1):
for j in range(cell_gradient_vector.shape[1] - 1):
block_vector = []
block_vector.extend(cell_gradient_vector[i][j])
block_vector.extend(cell_gradient_vector[i][j + 1])
block_vector.extend(cell_gradient_vector[i + 1][j])
block_vector.extend(cell_gradient_vector[i + 1][j + 1])
mag = lambda vector: math.sqrt(sum(i ** 2 for i in vector))
magnitude = mag(block_vector)
if magnitude != 0:
normalize = lambda block_vector, magnitude: [element / magnitude for element in block_vector]
block_vector = normalize(block_vector, magnitude)
hog_vector.append(block_vector)
return hog_vector, hog_image
def global_gradient(self):
gradient_values_x = cv2.Sobel(self.img, cv2.CV_64F, 1, 0, ksize=5)
gradient_values_y = cv2.Sobel(self.img, cv2.CV_64F, 0, 1, ksize=5)
gradient_magnitude = cv2.addWeighted(gradient_values_x, 0.5, gradient_values_y, 0.5, 0)
gradient_angle = cv2.phase(gradient_values_x, gradient_values_y, angleInDegrees=True)
return gradient_magnitude, gradient_angle
def cell_gradient(self, cell_magnitude, cell_angle):
orientation_centers = [0] * self.bin_size
for i in range(cell_magnitude.shape[0]):
for j in range(cell_magnitude.shape[1]):
gradient_strength = cell_magnitude[i][j]
gradient_angle = cell_angle[i][j]
min_angle, max_angle, mod = self.get_closest_bins(gradient_angle)
orientation_centers[min_angle] += (gradient_strength * (1 - (mod / self.angle_unit)))
orientation_centers[max_angle] += (gradient_strength * (mod / self.angle_unit))
return orientation_centers
def get_closest_bins(self, gradient_angle):
idx = int(gradient_angle / self.angle_unit)
mod = gradient_angle % self.angle_unit
if idx == self.bin_size:
return idx - 1, (idx) % self.bin_size, mod
return idx, (idx + 1) % self.bin_size, mod
def render_gradient(self, image, cell_gradient):
cell_width = self.cell_size / 2
max_mag = np.array(cell_gradient).max()
for x in range(cell_gradient.shape[0]):
for y in range(cell_gradient.shape[1]):
cell_grad = cell_gradient[x][y]
cell_grad /= max_mag
angle = 0
angle_gap = self.angle_unit
for magnitude in cell_grad:
angle_radian = math.radians(angle)
x1 = int(x * self.cell_size + magnitude * cell_width * math.cos(angle_radian))
y1 = int(y * self.cell_size + magnitude * cell_width * math.sin(angle_radian))
x2 = int(x * self.cell_size - magnitude * cell_width * math.cos(angle_radian))
y2 = int(y * self.cell_size - magnitude * cell_width * math.sin(angle_radian))
cv2.line(image, (y1, x1), (y2, x2), int(255 * math.sqrt(magnitude)))
angle += angle_gap
return image
img = cv2.imread('data/picture1.png', cv2.IMREAD_GRAYSCALE)
hog = Hog_descriptor(img, cell_size=8, bin_size=8)
vector, image = hog.extract()
plt.imshow(image, cmap=plt.cm.gray)
plt.show()HOG+SVM实现行人检测
数据集地址： ftp://ftp.inrialpes.fr/pub/lear/douze/data/INRIAPerson.ta
import cv2
import numpy as np
import random
def load_images(dirname, amout = 9999):
img_list = []
file = open(dirname)
img_name = file.readline()
while img_name != '': # 文件尾
img_name = dirname.rsplit(r'/', 1)[0] + r'/' + img_name.split('/', 1)[1].strip('\n')
img_list.append(cv2.imread(img_name))
img_name = file.readline()
amout -= 1
if amout <= 0: # 控制读取图片的数量
break
return img_list
# 从每一张没有人的原始图片中随机裁出10张64*128的图片作为负样本
def sample_neg(full_neg_lst, neg_list, size):
random.seed(1)
width, height = size[1], size[0]
for i in range(len(full_neg_lst)):
for j in range(10):
y = int(random.random() * (len(full_neg_lst[i]) - height))
x = int(random.random() * (len(full_neg_lst[i][0]) - width))
neg_list.append(full_neg_lst[i][y:y + height, x:x + width])
return neg_list
# wsize: 处理图片大小，通常64*128; 输入图片尺寸>= wsize
def computeHOGs(img_lst, gradient_lst, wsize=(128, 64)):
hog = cv2.HOGDescriptor()
# hog.winSize = wsize
for i in range(len(img_lst)):
if img_lst[i].shape[1] >= wsize[1] and img_lst[i].shape[0] >= wsize[0]:
roi = img_lst[i][(img_lst[i].shape[0] - wsize[0]) // 2: (img_lst[i].shape[0] - wsize[0]) // 2 + wsize[0], \
(img_lst[i].shape[1] - wsize[1]) // 2: (img_lst[i].shape[1] - wsize[1]) // 2 + wsize[1]]
gray = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)
gradient_lst.append(hog.compute(gray))
# return gradient_lst
def get_svm_detector(svm):
sv = svm.getSupportVectors()
rho, _, _ = svm.getDecisionFunction(0)
sv = np.transpose(sv)
return np.append(sv, [[-rho]], 0)
# 主程序
# 第一步：计算HOG特征
neg_list = []
pos_list = []
gradient_lst = []
labels = []
hard_neg_list = []
svm = cv2.ml.SVM_create()
pos_list = load_images(r'G:/python_project/INRIAPerson/96X160H96/Train/pos.lst')
full_neg_lst = load_images(r'G:/python_project/INRIAPerson/train_64x128_H96/neg.lst')
sample_neg(full_neg_lst, neg_list, [128, 64])
print(len(neg_list))
computeHOGs(pos_list, gradient_lst)
[labels.append(+1) for _ in range(len(pos_list))]
computeHOGs(neg_list, gradient_lst)
[labels.append(-1) for _ in range(len(neg_list))]
# 第二步：训练SVM
svm.setCoef0(0)
svm.setCoef0(0.0)
svm.setDegree(3)
criteria = (cv2.TERM_CRITERIA_MAX_ITER + cv2.TERM_CRITERIA_EPS, 1000, 1e-3)
svm.setTermCriteria(criteria)
svm.setGamma(0)
svm.setKernel(cv2.ml.SVM_LINEAR)
svm.setNu(0.5)
svm.setP(0.1) # for EPSILON_SVR, epsilon in loss function?
svm.setC(0.01) # From paper, soft classifier
svm.setType(cv2.ml.SVM_EPS_SVR) # C_SVC # EPSILON_SVR # may be also NU_SVR # do regression task
svm.train(np.array(gradient_lst), cv2.ml.ROW_SAMPLE, np.array(labels))
# 第三步：加入识别错误的样本，进行第二轮训练
# 参考 http://masikkk.com/article/SVM-HOG-HardExample/
hog = cv2.HOGDescriptor()
hard_neg_list.clear()
hog.setSVMDetector(get_svm_detector(svm))
for i in range(len(full_neg_lst)):
rects, wei = hog.detectMultiScale(full_neg_lst[i], winStride=(4, 4),padding=(8, 8), scale=1.05)
for (x,y,w,h) in rects:
hardExample = full_neg_lst[i][y:y+h, x:x+w]
hard_neg_list.append(cv2.resize(hardExample,(64,128)))
computeHOGs(hard_neg_list, gradient_lst)
[labels.append(-1) for _ in range(len(hard_neg_list))]
svm.train(np.array(gradient_lst), cv2.ml.ROW_SAMPLE, np.array(labels))
# 第四步：保存训练结果
hog.setSVMDetector(get_svm_detector(svm))
hog.save('myHogDector.bin')
test代码
import cv2
import numpy as np
hog = cv2.HOGDescriptor()
hog.load('myHogDector.bin')
cap = cv2.VideoCapture(0)
while True:
ok, img = cap.read()
rects, wei = hog.detectMultiScale(img, winStride=(4, 4),padding=(8, 8), scale=1.05)
for (x, y, w, h) in rects:
cv2.rectangle(img, (x, y), (x + w, y + h), (0, 0, 255), 2)
cv2.imshow('a', img)
if cv2.waitKey(1)&0xff == 27: # esc键
break
cv2.destroyAllWindows()代码参考：
https://github.com/PENGZhaoqing/Hog-feature/blob/master/hog.py
博客参考：
80行Python实现-HOG梯度特征提取_hog特征 python代码-CSDN博客
HOG特征 - 知乎
【计算机视觉】INRIA 行人数据集 （INRIA Person Dataset）_inria数据集-CSDN博客